1 / 12

F. López-Ostenero, V. Peinado, V. Sama & F. Verdejo

NE Recognition and Keyword-based Relevance Feedback in Mono and CL Automatic Speech Transcriptions Retrieval. F. López-Ostenero, V. Peinado, V. Sama & F. Verdejo. CLEF 2005. Vienna, Austria. September 22th 2005. 2. Goals. Test the suitability of our translation resources for a new track

pippa
Télécharger la présentation

F. López-Ostenero, V. Peinado, V. Sama & F. Verdejo

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. NE Recognition and Keyword-based Relevance Feedback in Mono and CL Automatic Speech Transcriptions Retrieval F. López-Ostenero, V. Peinado, V. Sama & F. Verdejo CLEF 2005. Vienna, Austria. September 22th 2005

  2. 2 Goals Test the suitability of our translation resources for a new track Test different strategies to clean the automatic transcriptions Check if Proper Noun Recognition can help to improve retrieval over automatic speech transcriptions Compare the efectiveness of manual and automatic keywords in a keyword-based pseudo-relevance feedback approach

  3. 3 Cleaning strategies Starting poing: ASR2004A field Join all the occurrences of a list of single characters: l i e b b a c h a r d  liebbachard Remove all extra occurences of duplicated words: “yes yes yes yes”  “yes” CLEAN collection Split documents of CLEAN collection in character 3-grams 3GRAMS collection Perform a morphological analysis over CLEAN collection and remove all words that cannot act as noun, adjective or verb MORPHO collection Perform a full Part of Speech tagging over CLEAN collection and delete all words except nouns, adjectives and verbs POS collection

  4. 4 Submitted runs Runs based on a query translation approach (using Pirkola’s structured queries) and INQUERY retrieval engine • Scores far from best monolingual and Spanish crosslingual runs: room for improvement • MAP scores of morpho and pos very similar, what’s the influence of cleaning strategies? • Character 3-grams scores worse than full word retrieval Only 5 different runs: not enough data to obtain clear conclusions

  5. 5 Proper noun identification We used a shallow entity recognizer to identify proper nouns in topics: • Monolingual: we identify proper nouns in English topics and we structure the query tagging them with a proximity operator • Crosslingual: we also identify proper nouns in Spanish topics but, which proper nouns should be translated and which ones should not? • if a proper noun appears in the SUMMARY field of documents, we assume that it should not be translated and we tag it using a proximity operator • otherwise we try to translate the proper noun La historia de Varian Fry y el Comité de Rescates de Emergencia … don’t translate try to translate

  6. 6 Pseudo Relevance Feedback Five collections to study a keyword-based pseudo relevance feedback: • AUTOKEYWORD2004A1 field (to build up AK1 collection) • AUTOKEYWORD2004A2 field (to build up AK2 collection) • Mix of 1 and 2 • single field keyword score, according to the position: 1st keyword = 20; 2nd keyword = 19 ... • if a keyword appears in both fields, its final score is the sum of both single field scores • select the top 20 scored keywords (to build up AK12 collection) • MANUALKEYWORD field (to build up MK collection) • Mix of 3 and 4 • Select the n manual keywords from 4 • If n < 20, add the 20 – n first keywords from 3 (to build up MKAK12 collection) Pseudo relevance feedback procedure: • Launch a plain query (without keyword expansion) • Retrieve keywords from top 10 retrieved documents • Mix these keywords using algorithm described to create AK12 collection • Expand query using top 20 keywords

  7. 7 Combining techniques Total combinations: 2 x 2 x 4 x 6 = 96 runs No 3grams in crosslingual runs: • mono: 2 x 4 x 6 = 48 runs • trans: 2 x 3 x 6 = 36 runs • real number or runs: 48 + 36 = 84 runs

  8. 8 Results Preliminary conclusions: • Monolingual improvement: 277.8% • Crosslingual improvement: 545.8% • Best strategies: • PRF using MK or MKAK12 collections • Use proper noun recognition • Monolingual 3-grams scores poorly, reaching a 27.2% of our best run

  9. 9 clean pos morpho Influence of proper nouns Each point represents MAP ent / MAP noent in percentage • monolingual: increment worthless and probably statistically not relevant • crosslingual: proper noun detection increment MAP more than twice

  10. 10 Influence of relevance feedback Each point represents MAP (rf method) / MAP (NO rf) in percentage • MK: the best option, but when combined with AK12 MAP decreases • AK12 usually better than AK1 or AK2. More stable in monolingual, better in crosslingual with entities and worse in crosslingual without entities • AK1 and AK2 identical in monolingual, AK1 better in crosslingual

  11. 11 Conclusions and future work The use of a shallow entity recognizer to identify proper nouns seems to be very useful, specially in a crosslingual environment where MAP increases 221,9% on average Cleaning methods based on full words (clean, morpho and pos) show no significative differences, but character 3-grams approach seems to be not useful for this task Pseudo Relevance Feedback using manually generated keywords shows to be the best option to improve the performance of the retrieval, with an average of 271.6% MAP regarding no relevance feedback Perform further analysis over the results, including statistical relevance tests Try a different approach to identify proper nouns in the automatic transcriptions or in the automatic keyword fields, instead of using the manual summary of the transcriptions

  12. NE Recognition and Keyword-based Relevance Feedback in Mono and CL Automatic Speech Transcriptions Retrieval F. López-Ostenero, V. Peinado, V. Sama & F. Verdejo CLEF 2005. Vienna, Austria. September 22th 2005

More Related